Documents & embeddings · Cobble Docs

Beyond chat models, Cobble hosts the pieces you need to make documents searchable and answerable: four OCR models and five embedding models. Both are billed from your wallet at per-page and per-token rates — they don't touch your plan window.

OCR: documents in, text out

Four models, one decision — accuracy versus cost:

Model	Batch size	Price	Pick it when
GLM-OCR	200 pages	$0.08/1K pages	Default for clean printed documents
DeepSeek OCR2	250 pages	$0.08/1K pages	Mixed layouts, tables
Nemotron OCR v2	300 pages	$0.08/1K pages	Large batches, throughput-first
Chandra OCR	500 pages	$4/1K pages	Hard documents: handwriting, degraded scans, complex layouts

Start with a $0.08 model; escalate pages that come back garbled to Chandra rather than running everything through it.

Embeddings: text in, vectors out

Model	Context	Price	Pick it when
`nomic/nomic-embed-1.5`	8K	$0.01/1M	Default — strong general retrieval, cheap
`qwen/qwen3-embedding-0.6b`	32K	$0.01/1M	Long chunks, multilingual
`qwen/qwen3-embedding-8b`	32K	$0.01/1M	Highest retrieval quality
`ibm/granite-embedding-97m`	8K	$0.09/1M	Latency-critical lookups
`ibm/granite-embedding-311m`	8K	$0.1/1M	Enterprise/compliance-flavored stacks

One rule that outranks model choice: embed your documents and your queries with the same model. Vectors from different models live in different spaces; mixing them silently breaks retrieval.

python

vectors = client.embeddings.create(
    model="nomic/nomic-embed-1.5",
    input=chunks,          # list of strings
).data

Recipe: PDF → searchable knowledge base

The full pipeline, using only Cobble endpoints plus a vector store of your choice:

python

# 1. OCR the document (wallet, per page)
text = cobble_ocr("contract.pdf", model="glm/glm-ocr")

# 2. Chunk — start simple: ~800 tokens per chunk, 100 overlap
chunks = chunk(text, size=800, overlap=100)

# 3. Embed the chunks (wallet, $0.01/1M tokens)
vectors = client.embeddings.create(model="nomic/nomic-embed-1.5", input=chunks)

# 4. Store in your vector DB (pgvector, Qdrant, Chroma...)
db.upsert(zip(chunks, vectors))

# 5. At question time: embed the query, retrieve, generate (plan window)
hits = db.search(embed(question), top_k=8)
answer = client.chat.completions.create(
    model="qwen/qwen3.5-122b-a10b",
    messages=[
        {"role": "system", "content": "Answer from the provided context. Cite which excerpt supports each claim."},
        {"role": "user", "content": f"Context:\n{format(hits)}\n\nQuestion: {question}"},
    ],
)

Cost intuition for a 1,000-page archive: ~$0.08 to OCR, roughly a cent to embed, and question-answering runs inside your plan window. The expensive part of RAG isn't the pipeline — it's doing it somewhere that meters every step.

Coming soon: a hosted reranker to slot between retrieval and generation, and managed knowledge endpoints that run this whole pipeline for you. The recipe above future-proofs cleanly — the reranker drops in between steps 5's retrieve and generate.